[fatbob]

Isn't this exactly the kind of thing that should be done by the compiler?

[KayEss]

Not really. The payoff is low unless it's done in exactly the right place. If I've already partitioned the work into threads the last thing I want is the compiler taking it upon itself to parallise loops that execute in those threads.

[ygra]

The compiler may do vectorization, but not splitting up things into multiple threads.